license: mit
base_model:
- dealignai/GLM-5.3-Flash-UNCENSORED-FP8
- zai-org/GLM-5.3-Flash
language: - en
- pt
library_name: exllamav3
pipeline_tag: text-generation
tags: - exl3
- exllamav3
- quantized
- 4-bit
- glm
- moe
- uncensored
- vision
- mtp
- long-context
GLM-5.3-Flash Uncensored · EXL3 4-bit · Bl4ck0ut
A 4-bit EXL3 quantization of dealignai/GLM-5.3-Flash-UNCENSORED-FP8,
calibrated on the Bl4ck0ut workload (long-context coding agents, tool calling, multi-file
reasoning) and measured against the original before publishing. Built to serve the full
1M-token context on a pair of 96 GB GPUs.
At a glance
| Architecture | Glm5NextForConditionalGeneration · 45 layers · 288 routed experts · MTP head · vision tower |
| Size | 163 GiB · 23 shards · 4.30 bpw overall |
| Experts | 4 bits (EXL3 trellis, mcg codebook) |
| Attention, KDA, indexer, shared experts, dense layers 0–2, lm_head | FP16, untouched |
| MTP head | 4 bits · Vision tower |
| Calibration | 503 rows × 2048 tokens of the bl4ck0ut workload |
| Engine | exllamav3 ≥ 1.4.5 (built with e648f1a; 1.4.6 recommended) |
| Context | 1,048,576 tokens |
Everything the model computes per token except the experts stays in FP16. That is the whole
recipe: compress the 288 experts, which are 95 % of the weights, and leave the path every token
walks through alone.
Fidelity
Measured on the published weights, greedy, against a frozen reference of the FP8 original.
| Test | Result |
|---|---|
| Held-out text (69,598 positions, 2k windows) | perplexity 1.7225 vs 1.7206 for the FP8 original (+0.1 %) · top-1 agreement 95.6 % · top-5 99.8 % |
| Passkey retrieval | correct at 8k, 128k, 512k and 1M tokens |
| 2000-token generation | no repetition loops, no degeneration |
| Expert routing (20k tokens) | no layer with more than 2 never-selected experts; no expert above 10.5 % of picks |
| MTP (speculative) head | output identical to non-speculative decoding · 85.5 % draft acceptance |
| Tool calling | valid <tool_call> with the model's own template |
| Vision | working at 6 bits |
Every expert's input scale was audited from the shards (manifest.json in this repo): 37,152
expert tensors, none outside the sane range, one benign outlier.
Running it
Any exllamav3-based server works (TabbyAPI, the exllamav3 examples). Two 96 GB GPUs hold the
model and a 1M-token cache; four make prefill at 1M take about eight minutes.
- Keep
generation_config.jsonandchat_template.jinjaas shipped. They are the ones from
the upstream revision the reference was frozen on, so measurements stay comparable. - Leave speculative decoding (MTP) off for now: the head is accurate, but the current engine
implementation makes it slower, not faster. - This is an uncensored model. It follows instructions it would otherwise refuse. Use it
where that is what you want, and nowhere else.
Provenance
- Base: zai-org/GLM-5.3-Flash (MIT).
- Uncensored FP8: dealignai/GLM-5.3-Flash-UNCENSORED-FP8, revision
d21b1956. - Quantization: Bl4ck0ut, September 2026. FP8 → BF16 → EXL3 on 4× RTX PRO 6000, with three
fixes to the multi-GPU path of the converter (stream synchronization in the scale search,
per-module dispatch, device selection for small tensors); the patches ship with the pipeline and are being sent upstream.
Made by Bl4ck0ut.